Every demo below runs against the v2 server. The scripts exit 0 only when
all their checks pass, and the three under scripts/ are also run by
tests/demo_smoke.rs on every test run, which asserts their headline
results.
| Demo | Run | What it shows |
|---|---|---|
| Watch it learn | cargo run --release --bin syntra -- demo |
A server on a fresh temporary store with simulated LLM-routing traffic (made-up quality and cost per route and task). Prints the admin console's address and key, and every ten seconds the share of the best possible expected reward the capsule earns. Ctrl-C stops it. |
| LLM routing | examples/llm-routing | syntra.llm.ModelRouter routing 4,000 requests between three simulated model routes with delayed quality grades, decisions made in-process and verified by the server, learning per task, then syntra evaluate --store on the logs: DR estimates of the constant policies land on the simulation's known values, and a gate refuses always-large. Needs the Python extension (sdk/python/scripts/develop.sh). |
| Containment matrix | scripts/demo-containment.py | 14 attack vectors from a feature program wired to every I/O capability: file and symlink escapes toward a canary outside the sandbox, a policy that tries to widen its file root, SSRF to metadata, RFC1918 and the admin console, plain http to an allowed host, exfiltration, policy flips, a compute-budget abort. 24/24 checks; every denial is a 500 audited as execution_denied with its request id. |
| TLS gateway | scripts/demo-tls-gateway.py | The server behind a TLS reverse proxy: spec, program install, decide and reward over TLS, wrong CA and hostname mismatch rejected, plain HTTP refused. 8/8 checks. |
| Agent governor | scripts/demo-agent-governor.py | Agent tool calls decided by a learning capsule: a budget rail in the feature program (exclude.<id>) that forces block with probability 1, learned allow/hold per agent from rewards, byte-identical model after a restart, tenant isolation. 10/10 checks. |
scripts/smoke-test.sh is a quicker check of a build: 32 checks of auth, decide and reward, idempotency, scoped tokens, a feature program and its sandbox, metrics, evaluation, restart persistence, backup and doctor, in about a second.
The v1 demos (LLM model routing and governed routing with a replay gate,
both shell scripts, and anomaly-aware routing) went away with the v1
API. Model routing with delayed feedback is now examples/llm-routing,
the replay gate is syntra evaluate and POST .../promote, and
computing signals before a decision is a feature program, as in the
containment and agent-governor demos. The science demos (Mars transfer
search, edge of chaos, pandemic policy scoring and others), the proof lab
and self-evolving capsules live in the separate Lycan Lab repository.